EFFICIENT LEARNING-BASED DIGITAL PRE-DISTORTION AND MULTI-IMPAIRMENT COMPENSATION IN MIMO TRANSMITTERS

Abstract

Minimizing energy consumption in cellular base station power amplifiers (PAs) is crucial, as they contribute to 50-80% of total power use. Operating PAs more efficiently near saturation introduces nonlinear distortion, addressed by digital predistortion (DPD). However, DPD faces challenges, including high PA nonlinearity and TX architecture impairments. Neural networks (NNs) offer improvements but add complexity. The first section of this thesis highlights the optimization of NN model complexity through strategic input layer design, employing envelope-dependent terms and residual learning. Simulation and experimental results demonstrate superior performance over conventional models, showcasing a reduced complexity of 26.40%. The second part introduces an integrated DPD solution, merging ILC architecture with NN models to mitigate impairments in direct conversion TXs (DCTs). It exhibits superior in-band and out-of-band performance compared to existing DPD models. The third section introduces low-complexity beam-oriented DPD (BO-DPD) techniques for mMIMO TXs. These techniques achieve comparable performance to conventional multi-DPD approaches with significantly reduced complexity.